Skip to main content

A Python package for benchmarking adversarial attacks and defenses.

Project description

AttackBenchLib: Evaluating Gradient-based Attacks for Adversarial Examples

Riccardo Trebiani, Antonio Emanuele Cinà, Jérôme Rony, Maura Pintor, Luca Demetrio, Ambra Demontis, Battista Biggio, Ismail Ben Ayed and Fabio Roli

Leaderboard: https://attackbench.github.io/

Paper: https://arxiv.org/pdf/2404.19460

Tutorial Notebook: Open In Colab

How it works

AttackBenchLib is a library that implements the framework described in the AttackBench paper in a new modular, user-friendly way in order to make multiple workflows and kinds of analysis possible through the use of a single library. The AttackBench framework aims to fairly compare gradient-based attacks based on their security evaluation curves. To this end, we derive a process involving five distinct stages, as depicted below.

  • In stage (1), we construct a list of diverse non-robust and robust models to assess the attacks' impact on various settings, thus testing their adaptability to diverse defensive strategies.
  • In stage (2), we define an environment for testing gradient-based attacks under a systematic and reproducible protocol. This step provides common ground with shared assumptions, advantages, and limitations. We then run the attacks against the selected models individually and collect the performance metrics of interest in our analysis, which are perturbation size, execution time, and query usage.
  • In stage (3), we gather all the previously-obtained results, comparing attacks with the novel local optimality metric.
  • Finally, in stage (4), we aggregate the optimality results from all considered models, and in stage (5) we rank the attacks based on their average optimality, namely global optimality.

Currently implemented

Attack Original Advertorch Adv_lib ART CleverHans DeepRobust Foolbox Torchattacks
DDN
ALMA
FMN
PGD
JSMA
CW-L2 ~
CW-LINF
FGSM
BB
DF ~
SuperDF
APGD
BIM
EAD
PDGD
PDPGD
TR
FAB

Legend:

  • empty : not implemented yet
  • ☒ : not available
  • ✓ : implemented
  • ~ : not functional yet

Requirements and Installation

  • Python >= 3.9, < 3.13
  • PyTorch >= 2.4
  • TorchVision >= 0.19
  • CUDA compatible GPU (recommended)

Install from PyPI

pip install attackbenchlib

Optional dependencies

# Attack library wrappers (ART, Foolbox, Torchattacks, CleverHans)
pip install "attackbenchlib[attacks]"

# Model loading utilities (RobustBench, timm, transformers)
pip install "attackbenchlib[models]"

# Analysis and visualization tools (scikit-learn, seaborn, plotly)
pip install "attackbenchlib[metrics]"

# Everything (attacks + models + metrics)
pip install "attackbenchlib[all]"

Note on autoattack: RobustBench depends on autoattack. If you encounter import errors related to autoattack after installing attackbenchlib[models], install it manually from GitHub:

pip install git+https://github.com/fra31/auto-attack

Note on adv-lib: The Adversarial Library (adv-lib) is not available on PyPI. If you need adv-lib attacks, install it manually:

pip install git+https://github.com/jeromerony/adversarial-library

Note on deeprobust: Requires scipy<1.8.0 and only works on Python 3.9: pip install "attackbenchlib[deeprobust]"

Google Colab

On Google Colab, install with all dependencies:

!pip install "attackbenchlib[models,attacks]" -q
!pip install git+https://github.com/fra31/auto-attack -q  # required for RobustBench

You may see red dependency conflict warnings during installation. These are caused by RobustBench's strict dependency pins (e.g., timm==1.0.9) conflicting with Colab's pre-installed packages. They are harmless warnings — the library works correctly.

Install from source (development)

git clone https://github.com/attackbench/AttackBenchLib.git
cd AttackBenchLib
pip install -e ".[dev]"

Usage

import torch
import attackbench
from attackbench.attacks import apgd

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Load model and dataset (requires attackbenchlib[models])
model = attackbench.load_model('Standard', dataset='cifar10', threat_model='Linf')
model.to(device)

dataset = attackbench.get_loader(dataset='cifar10', batch_size=128, num_samples=1000)

# Run attack
results = attackbench.run_attack(
    model=model,
    dataset=dataset,
    attack=apgd,
    threat_model='linf',
    device=device
)

# Analyze results (requires attackbenchlib[metrics])
stats = attackbench.get_stats(results, 'linf')
print(f"ASR: {stats['ASR']*100:.1f}%")

Preconfigured attacks available out of the box: pgd, fgsm, apgd, fab, fmn, deepfool, superdeepfool, trust_region.

To use attacks from external libraries (requires attackbenchlib[attacks]):

# List available attacks
attacks = attackbench.list_attacks(threat_model='linf')

# Load a specific library attack
art_pgd = attackbench.get_attack(lib='art', attack='pgd', threat_model='linf')
results = attackbench.run_attack(model=model, dataset=dataset, attack=art_pgd, threat_model='linf', device=device)

Attack format

The wrappers for all the implementations (including libraries) must have the following format:

  • inputs:
    • model: nn.Module taking inputs in the [0, 1] range and returning logits in $\mathbb{R}^K$
    • inputs: FloatTensor representing the input samples in the [0, 1] range
    • labels: LongTensor representing the labels of the samples
    • targets: LongTensor or None representing the targets associated to each samples
    • targeted: bool flag indicating if a targeted attack should be performed
  • output:
    • adv_inputs: FloatTensor representing the perturbed inputs in the [0, 1] range

Citation

If you use the AttackBench leaderboards or implementation, then consider citing our paper:

@inproceedings{cina2025attackbench,
  title={Attackbench: Evaluating gradient-based attacks for adversarial examples},
  author={Cin{\`a}, Antonio Emanuele and Rony, J{\'e}r{\^o}me and Pintor, Maura and Demetrio, Luca and Demontis, Ambra and Biggio, Battista and Ayed, Ismail Ben and Roli, Fabio},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={3},
  pages={2600--2608},
  year={2025},
  DOI={10.1609/aaai.v39i3.32263}
}

Contact

Feel free to contact us about anything related to AttackBench by creating an issue, a pull request or by email at antonio.cina@unige.it.

Acknowledgements

AttackBench has been partially developed with the support of European Union’s ELSA – European Lighthouse on Secure and Safe AI, Horizon Europe, grant agreement No. 101070617, and Sec4AI4Sec - Cybersecurity for AI-Augmented Systems, Horizon Europe, grant agreement No. 101120393.

sec4ai4sec    elsa    europe

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

attackbenchlib-1.0.9.tar.gz (478.1 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

attackbenchlib-1.0.9-py3-none-any.whl (146.2 kB view details)

Uploaded Python 3

File details

Details for the file attackbenchlib-1.0.9.tar.gz.

File metadata

  • Download URL: attackbenchlib-1.0.9.tar.gz
  • Upload date:
  • Size: 478.1 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.3

File hashes

Hashes for attackbenchlib-1.0.9.tar.gz
Algorithm Hash digest
SHA256 ea39ecb29228c4440250c275d29925fdd0ef2d123df720758f0b203db89dd7a7
MD5 d6ffcb0eb4218a7336b4439b7362f448
BLAKE2b-256 aa9703379975d59c73c5bdef6efc4d06f0c122e89df0bd9b9fcd9de4e92c0674

See more details on using hashes here.

File details

Details for the file attackbenchlib-1.0.9-py3-none-any.whl.

File metadata

  • Download URL: attackbenchlib-1.0.9-py3-none-any.whl
  • Upload date:
  • Size: 146.2 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.12.3

File hashes

Hashes for attackbenchlib-1.0.9-py3-none-any.whl
Algorithm Hash digest
SHA256 c319f4e24217f57b0a8e8ca3417e2aef239e89ba0ac588f63996e569824445ba
MD5 bdd72874e7b400af51ed11b8a374a383
BLAKE2b-256 a5a21806f15bea58b04f9e948286603dce157c459587d3cc5e883ece259fcb06

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page